RunningTab — tracking unfinished workspace requirements

TL;DR

RunningTab keeps requirements and reading evidence in the environment, helping an agent turn workspace files into a complete deliverable. (source)

Authors & Org

Jinheon Baek, Soyeong Jeong, Yumin Choi, Dongsu Han and Sung Ju Hwang. The abstract record does not establish affiliations. (source)

Method

The agent declares requirements. The environment records file reads as excerpts with provenance and listed-but-unopened files as candidates. Requirements are displayed beside matching excerpts and candidates; the agent resolves them or records why they are set aside. A finish check returns requirements still open. (source)

Results

The authors report consistent improvements over plain direct workspace interaction and model-maintained records across three benchmarks and three LLMs. The abstract gives no scores, model names, benchmark names or cost figures. (source)

Significance

For Agents (LLM Agents), this separates access to evidence from tracking whether the task's requirements were satisfied. It targets omissions that can persist even after the agent reads the necessary figure. (source)

Open Questions

How much overhead does the record add, and how reliable is requirement resolution? The abstract alone does not answer these questions; the full paper was not read. (source)

Cite

• arXiv:2610.10444, submitted 2026-10-07.
• Captured abstract record.